Analyze Kernel Bottleneck

Systematically identify whether GPU kernel is compute-bound, memory-bound, or latency-bound. Measure baseline performance. Classify on roofline. Compute occupancy and compute/load ratio per tile. Inspect SASS instruction mix and stall codes. Check shared memory cliff. Apply decision matrix to select right optimization strategy.

When Use

• Before optimizing any CUDA kernel -- establish baseline and classify bottleneck type
• After writing first working version of kernel to identify optimization path
• Kernel underperforms expectations relative to theoretical peak
• Deciding between cp.async, larger tiles, or algorithmic restructuring

Inputs

• Required: Compiled kernel (.cubin or .cu source with build command)
• Required: Benchmark harness that launches kernel with CUDA event timing
• Required: Problem dimensions (e.g., M, N, K for GEMM; seq_len, heads, head_dim for attention)
• Optional: Target GPU architecture (default: GA104 / sm_86 / RTX 3070 Ti)
• Optional: Expected peak utilization percentage for comparison
• Optional: Prior profiling data (Nsight Compute reports)

Steps

Step 1: Measure Baseline Performance

Run kernel with CUDA events (BenchTimer). Record time in milliseconds. Calculate effective throughput metrics:

1. Compile kernel if not already built:

   nvcc --cubin -arch=sm_86 -O2 -o kernel.sm_86.cubin kernel.cu
   nvcc -arch=sm_86 -O2 -o bench bench.cu -lcuda -I../../phase2/common
   

2. Run with representative problem sizes, ensuring warmup runs precede measurement:

   ./bench 4096 4096 4096
   

3. Record kernel time in ms from CUDA events (not wall-clock).
4. Calculate effective GFLOPS and effective bandwidth:
   • GEMM: effective_gflops = (2 * M * N * K) / (time_ms / 1000) / 1e9
   • Bandwidth-limited kernels: effective_bw = total_bytes / (time_ms / 1000) / 1e9
   • Flash Attention: effective_gflops = (4 * batch * heads * seq_len^2 * head_dim) / (time_ms / 1000) / 1e9

Got: Baseline numbers: kernel time in ms, effective GFLOPS, effective bandwidth.

If fail: Check kernel launches without error (CHECK_CU macro). Verify warmup runs precede measurement. Ensure problem dimensions large enough to saturate GPU (small problems may bottleneck on launch overhead).

Step 2: Classify on Roofline

Compute arithmetic intensity. Compare against machine balance point to classify kernel:

1. Calculate arithmetic intensity: AI = FLOPs / bytes_loaded_from_global_memory. Count only unique bytes loaded from DRAM (not shared memory or register reuse).
2. Look up machine balance point: balance = peak_compute / peak_bandwidth.
3. Classify: AI < balance? Kernel memory-bound. AI > balance? Kernel compute-bound.

GA104 (RTX 3070 Ti) Reference Values:

Resource	Peak	Unit
FP32 FFMA	21.7	TFLOPS
FP16 Tensor Core (HMMA)	174	TFLOPS
INT8 Tensor Core (IMMA)	696	TOPS
DRAM Bandwidth	608	GB/s
L2 Cache	4	MB
SMs	48

Derived Balance Points:

Precision	Balance Point (FLOP/byte)
FP32 FFMA	21700 / 608 = 35.7
FP16 TC	174000 / 608 = 286.2
INT8 TC	696000 / 608 = 1144.7

4. Compute attained fraction: attained = effective_throughput / peak_throughput. Memory-bound? Compare effective bandwidth to 608 GB/s. Compute-bound? Compare effective GFLOPS to relevant peak.

Got: Classification as compute-bound, memory-bound, or latency-bound (low occupancy causing neither compute nor memory saturation) with numerical justification.

If fail: Recheck byte counting. Watch for redundant re-reads (e.g., 9x in direct conv2d without im2col). Neither compute nor memory saturated? Kernel likely latency-bound (see Step 3).

Step 3: Calculate Occupancy

Determine active warps per SM from launch configuration and resource usage:

1. Extract resource usage:

   nvcc --cubin -arch=sm_86 -O2 --resource-usage -o kernel.sm_86.cubin kernel.cu 2>&1 | grep -E 'registers|smem'
   

2. From launch config: warps_per_block = threads_per_block / 32.
3. Compute blocks/SM from each limiting factor:
   • Register limit: floor(65536 / (registers_per_thread * threads_per_block))
   • Smem limit: floor(available_smem_per_SM / smem_per_block) -- see Step 6 for cliff
   • Warp limit: floor(48 / warps_per_block) (GA104 max: 48 warps/SM)
   • Block limit: 16 blocks/SM max on GA104
4. Actual blocks/SM = min(register_limit, smem_limit, warp_limit, block_limit).
5. Active warps/SM = blocks_per_SM * warps_per_block.
6. Key threshold: 8 warps/SM sufficient for latency hiding on GA104. Below 8 = structural problem causing latency-bound behavior.

Got: Occupancy table showing blocks/SM, active warps/SM, limiting factor (registers, smem, or warps).

If fail: Check cuFuncSetAttribute for dynamic shared memory. Verify --resource-usage reports match actual launch configuration. Register count unexpectedly high? Try --maxrregcount=N to cap registers (trading register spills for occupancy).

Step 4: Compute Compute/Load Ratio Per Tile

Count compute instructions and load bytes per K-tile from SASS (not source code):

1. Disassemble:

   cuobjdump -sass kernel.sm_86.cubin > kernel.sass
   

2. Count compute instructions per tile (inner loop over one K-tile):
   • grep -c 'HMMA' kernel.sass -- FP16 Tensor Core ops
   • grep -c 'IMMA' kernel.sass -- INT8 Tensor Core ops
   • grep -c 'FFMA' kernel.sass -- FP32 fused multiply-add
3. Count global loads per tile:
   • grep -c 'LDG' kernel.sass -- global memory loads
   • Multiply by bytes per load (typically 16 bytes for LDG.128)
4. Calculate ratio: compute_ops / load_ops per tile.
5. Classify using cp.async decision threshold (from gpu_reflections.md Insight 2):
   • High (>20:1): cp.async net-negative; warp interleaving already hides DRAM latency. Focus on algorithmic changes. Reference: Flash Attention has 64 HMMA per tile = high ratio, cp.async measured -5%.
   • Medium (5-20:1): cp.async may help, benchmark both paths.
   • Low (<5:1): cp.async strongly beneficial; loads dominate and async copy hides latency. Reference: IGEMM has 8 IMMA per tile = low ratio, cp.async measured +35%.

Got: Compute/load ratio with classification (high/medium/low) and cp.async recommendation.

If fail: Count from SASS disassembly, not source code -- compiler may fuse, eliminate, or reorder instructions. Ensure counting instructions within inner loop only (K-tile iteration), not entire kernel.

Step 5: Inspect SASS Instructions

Examine full SASS instruction mix and stall codes:

1. Disassemble (if not done in Step 4):

   cuobjdump -sass kernel.sm_86.cubin > kernel.sass
   

2. Count key instruction types:

   grep -c 'HMMA.16816' kernel.sass      # FP16 Tensor Core
   grep -c 'IMMA.16816' kernel.sass      # INT8 Tensor Core
   grep -c 'FFMA' kernel.sass            # FP32 fused multiply-add
   grep -c 'LDGSTS' kernel.sass          # cp.async (global->shared)
   grep -c 'LDG' kernel.sass             # Global load
   grep -c 'STS' kernel.sass             # Shared store
   grep -c 'LDS' kernel.sass             # Shared load
   grep -c 'BAR.SYNC' kernel.sass        # Barrier synchronization
   grep -c 'SHFL' kernel.sass            # Warp shuffle (reductions)
   grep -c 'MUFU' kernel.sass            # Special function unit
   

3. Check stall codes on critical instructions:

   grep 'HMMA' kernel.sass | head -5     # Expect S08 minimum (hardware constraint)
   grep 'IMMA' kernel.sass | head -5     # Compiler emits S04, reducible to S02 via CuAssembler
   grep 'FFMA' kernel.sass | head -5     # Check for S04 (reducible to S01 on independent FFMAs)
   

4. Identify optimization targets:
   • HMMA S08 stalls: hardware minimum on Ampere, cannot be reduced. Focus elsewhere.
   • IMMA S04 stalls: compiler conservative. CuAssembler can tighten to S02 (measured 15-20% gain).
   • FFMA S04 stalls: if independent, reducible to S01 via CuAssembler.
   • Excessive BAR.SYNC: may indicate over-synchronization between pipeline stages.

Got: Instruction count table and stall code summary with identified optimization targets.

If fail: Ensure cuobjdump architecture matches kernel compilation target (both must be sm_86). SASS output empty? Cubin may be corrupt -- recompile.

Step 6: Check Smem Cliff

Determine whether shared memory usage crosses architecture-specific occupancy cliff:

1. Read smem/block from --resource-usage output (Step 3) or cuobjdump --res-usage kernel.sm_86.cubin.
2. Compare against cliff threshold:
   • GA104 (sm_86): 100 KB max smem/SM. Cliff at 50 KB/block.
   • Confirmed empirically: 48 KB/block -> 2 blocks/SM (good), 56 KB/block -> 1 block/SM (2x regression).
3. Above cliff (smem > 50 KB/block):
   • Blocks/SM drops to 1, active warps drop to warps_per_block (typically 4).
   • 2x performance regression expected from exposed DRAM stalls.
4. Check double-buffering impact: Double-buffering doubles smem usage. Current smem 30 KB? Double-buffered = 60 KB, crosses cliff. Evaluate whether async benefit outweighs occupancy loss.
5. Record smem/block, blocks/SM, and whether cliff crossed.

Got: Smem/block value with blocks/SM count and explicit statement of whether 50 KB cliff crossed.

If fail: Above cliff and occupancy is bottleneck? Optimization strategy must change: reduce tile size to get smem under 50 KB, or accept 1 block/SM and compensate with higher compute/load ratio per tile (more register reuse, longer K-tiles).

Step 7: Build Decision Matrix

Synthesize findings from Steps 2-6 into optimization strategy:

Condition	Strategy
Memory-bound + low compute/load ratio (<5:1) + smem under cliff	Software pipelining with cp.async (LDGSTS). Overlap global loads with compute.
Memory-bound + high compute/load ratio (>20:1) + 8+ warps	Warp interleaving already hides latency. Focus on algorithmic changes: implicit GEMM, split-Q, im2col.
Compute-bound + FFMA-heavy	CuAssembler stall code tightening: S04 -> S01 on independent FFMAs.
Compute-bound + HMMA-heavy	S08 is hardware minimum, cannot reduce. Increase tile reuse (larger M/N tiles, longer K-loop).
Compute-bound + IMMA-heavy	CuAssembler: S04 -> S02 on IMMA instructions (compiler is conservative).
Latency-bound (low occupancy, neither saturated)	Reduce smem or registers to get more blocks/SM. Get above 8 warps/SM.
Smem above cliff	Reduce tile size or restructure to get smem/block under 50 KB (GA104).

1. Rank applicable strategies by expected gain, using compute/load ratio and occupancy data.
2. Estimate gain range for each strategy based on how far kernel is from relevant ceiling.
3. Flag conflicts: e.g., cp.async doubles smem (may cross cliff), larger tiles increase register pressure (may reduce occupancy).

Got: Ranked list of recommended optimizations with predicted gain range and potential conflicts.

If fail: No clear winner emerges? Run micro-benchmarks isolating each strategy (e.g., test cp.async alone, test reduced tile size alone) to measure actual impact before combining.

Step 8: Document Findings

Produce structured bottleneck report:

1. Baseline: kernel time, effective GFLOPS, effective bandwidth, problem dimensions.
2. Roofline position: arithmetic intensity, classification, attained fraction of peak.
3. Occupancy: blocks/SM, active warps/SM, limiting factor.
4. Compute/load ratio: ratio value, classification (high/medium/low), cp.async recommendation.
5. SASS summary: instruction counts table, stall code findings, CuAssembler targets.
6. Smem cliff: smem/block, blocks/SM, cliff status.
7. Recommendation: ranked optimization strategies with gain estimates.

## Bottleneck Analysis Report: [kernel_name]

### Baseline
- Problem: [dimensions]
- Kernel time: [X] ms
- Effective GFLOPS: [Y] | Effective BW: [Z] GB/s

### Roofline Classification
- Arithmetic intensity: [AI] FLOP/byte
- Balance point: [BP] FLOP/byte ([precision])
- Classification: **[compute|memory|latency]-bound**
- Attained fraction: [X]% of peak

### Occupancy
| Resource | Per Block | Limit/SM | Blocks/SM |
|----------|-----------|----------|-----------|
| Registers | [N]/thread | 65536 | [B] |
| Shared mem | [X] KB | 100 KB (cliff: 50 KB) | [B] |
| Warps | [W] | 48 | [B] |
| **Limiting** | | | **[min(B)]** |
- Active warps/SM: [W] ([sufficient|insufficient] for latency hiding)

### Compute/Load Ratio
- Compute ops/tile: [N] [HMMA|IMMA|FFMA]
- Load bytes/tile: [N] bytes ([N] LDG x [N] bytes)
- Ratio: [X]:1 — **[high|medium|low]**
- cp.async recommendation: [beneficial|neutral|detrimental]

### SASS Instruction Mix
| Instruction | Count | Notes |
|-------------|-------|-------|
| HMMA.16816 | [N] | Stall: S08 (hardware min) |
| IMMA.16816 | [N] | Stall: S04 (reducible to S02) |
| FFMA | [N] | Stall: S04 (reducible to S01) |
| LDG | [N] | |
| LDGSTS | [N] | cp.async |
| BAR.SYNC | [N] | |

### Smem Cliff
- Smem/block: [X] KB — [under|over] 50 KB cliff
- Blocks/SM: [B] — [no occupancy loss|occupancy halved]

### Recommended Optimizations (ranked)
1. [Strategy] — estimated [X-Y]% gain
2. [Strategy] — estimated [X-Y]% gain
3. [Strategy] — estimated [X-Y]% gain

Got: Complete markdown report consumable by kernel-optimizer agent or human developer.

If fail: Re-run with different problem sizes (e.g., 1024, 2048, 4096, 8192) to confirm findings not size-specific. Small problems may appear latency-bound when real bottleneck at scale is memory bandwidth.

Checks

• Baseline measured with CUDA events (not wall-clock)
• Roofline classification determined (compute/memory/latency bound)
• Occupancy computed with limiting factor identified
• Compute/load ratio per tile calculated from SASS
• SASS instruction mix and stall codes documented
• Smem cliff checked against architecture threshold
• Decision matrix applied with strategy recommendation
• Findings documented in structured report

Pitfalls

• Re-read multiplication: Direct conv2d reads each weight 9x without im2col. Inflates byte count by 9x. Use actual unique bytes loaded from DRAM, not total load instructions, when computing arithmetic intensity.
• Confusing FP16 Tensor Core peak with FP32 peak: FP16 TC peak 174 TFLOPS, FP32 FFMA peak 21.7 TFLOPS -- 8x difference. Wrong peak makes roofline classification meaningless.
• Using 64 KB as smem cliff instead of 50 KB on GA104: GA104 (sm_86) has 100 KB max smem/SM. Cliff at 100/2 = 50 KB/block, not 64 KB. Architecture-specific; other GPUs differ.
• Ignoring warp interleaving when evaluating cp.async: 8 warps with long compute phases (high compute/load ratio) already hide DRAM latency through warp scheduling. Adding cp.async in this regime adds smem pressure and barrier overhead for no benefit (measured -5% on Flash Attention).
• Counting instructions from source code instead of SASS: Compiler may fuse operations, eliminate dead code, unroll loops differently, or reorder instructions. Always count from cuobjdump -sass output.
• Not running warmup iterations: First kernel launch includes JIT compilation overhead and cold cache effects. Always run 2-5 warmup iterations before measured run.

See Also

• pipeline-gpu-kernel -- implement software pipelining with cp.async when analysis identifies memory-bound kernel with low compute/load ratio
• simulate-cpu-architecture -- complementary architecture analysis for CPU-side bottlenecks in host-device workflows